RESOURCE DIMENSIONING ASPECT OF HETEROGENEOUS TRAFFIC WITH DIFFERENT SERVICE REQlUIREMENTS: INTEGRATION VERSUS SEGREGATION'

In this paper, we cansider the ATM networks in which the Virtual Path concept is implemented. The question of how to multiplex two or more diverse traflic classes while providing different Quality of Service requirements is a very complicated, open problem. Two distint options are available: integration and segregation. In an integration approach al l the W c from Merent c o ~ e c t i ~ ~ l s are multiplexed onto one VP. This implies that the most restrictive QOS requirements must be applied to all services. Therefore, link utilization will be decreased because unnecessarily stringent QOS is provided to al l C O M ~ O I I S . With the segregation approach the problem can be much simplified if different types of traflic are separated by assigning a VP with dedicated resources (bUgers and links). Therefore, resources may not be aciently utilized because no sharing of bandwidth can take place across the VP. The probability that the bandwidth required by the accepted co~ections exceeds the capacity of the link is evaluated with the Probability of Congestion (PC). Since the PC can be expressed as the CLP, we shall simply cany out bandwidth allocation using the PC. We first focus on the influence of some parameters (CLP, bit rate and burstiness) on the capacity required by a VP supporting a single traffic class using the New Convolution Approach. Numerical resuits are presented both to compare the required capacity and to observe which conditions under each approach are preferred.


INTRODUCTION
In Asynchronous 'Transfer Mode (ATM) networks, different traffic classes with different flow characteristics and Qualrty of ScMce (QOS) requirements are stabtically multiplexed to increase the utilisation of link bandwidth.
QOS is normally expressed in terms of Cell Loss Probability (CLP), cell delay and jitter, in which both cell delay and jitter c a l f easily be imposed by having a f i t e buffer to limit the maxi" queue length.Therefore, we consider CLP requirements as QOS.
If statistical multiplexing is considered, the bandwidth required by the accepted connections could exceed the capacity of the hlk.The probability that the bandwidth required by the accepted ~~~e c t i o n s exceeds the capacity of the link is evaluated by means of the Probability of Congestion (PC).Thus, PC, Total CLP and Individual CLP ca13 be calculated by using the New Convolution Approach (NCA).
In this paper, we mnsider those ATM networks in which the Virtual Path (VP) concept is implemented.How to multiplex two or more diverse t r a c classes while providmg different QOS requirements is a very complicated, open problem.Two distinct options are available: integration and segregation.
The paper is organised as follows.Section 2 describe the segregation and integration approach to virtual traffic management.In section 3 we present an algorithm to determine the minimum an sufficient capacity required by both approaches so that all the individual QOS requirements are guaranteed.In section 4 we explain how the New Convolution Approach (NCA) computes the Individual Cell Loss Probabilities and how we can compute the Required Capacity in this case.Section 5 and 6 deal with numerical results , followed by conclusions and further works in Section 7.

INTEGRATION VS. SEGREGATION
When two traftic flows with different QOS requirements (i.e.: Merent Cell Loss Probabihties) arrive to a node and, these share the same destination node, as far as local decisions are concerned, two possible and simple approaches to virtual traffic management can be taken [I], [2], [31, [4], [5]: In an INTEGRATION APPROACH3 (Sharing) all the traffic from different connections are multiplexed onto one VP.While sharing the available bandwidth and buffers.The capacity allocated to the VP should be large enough to satisfy all the individual QOS.Arriving cells from all source types are routed to a single FCFS buffer.In absence of buffer priorities, the same CLP is provided to both ImfEc types.This implies that the most restrictive QOS requirements must be applied to all services.Therefore, link utilisation will be decrease because unnecessarily stringent QOS is provided to all connections and efficient use of available resources isdiflicdt since most source types receive a QOS that is much better than the one they had demanded.This Eact ' could produce lower multiplexing gain than the segregation approach.The multiplexing gain is strongly dependent on the ratio between total link bandwidth and source mean bandwidth, concentrating a greater traf3ic intensity could conceivably produce a higher gain than segregation, even with the most stringent QOS requirements.
. FCFS buffers, which are then served in accordance to a chosen bandwidth assignment policy.Therefore, the resources may not be efficiently uthsed because no sharing of bandwi& can take place across the VP.In this case, advantage can be taken on the less stringent QOS to achieve a high multiplexing gain.As a consequence of the separation, more than one VP may be established between the same sourcedestination pair, with each Carrying different types of traffic.The capacity allocated to each VP should be sufficient to guarantee the required QOS of the traffic classes supported.The sum of all the assigned VP capacities would thus constitute the total capacity required.aggregation by some authors.

ALGORITHM TO DETERMINE THE MlNIMUM CAPACITY REQUIRED BY A VIRTUAL
PATH Superior performance of one or the other strategy depends on the particular characteristics of the involved traffic flows.There is a need of algorithms that, on the basis of small set of easily estimated parameters for each traffic flow, can predict accurately the benefits of aggregation / segregation.
In [6] we propose the New Convolution Approach to calculate the bandwidth required by dlfferent traf€ic types to achieve a certain QOS.When using this results, an iterative algorithm determines the sufficient and minimum capacity of a VP so that all the individual QOS are satisfied whenever different traflic classes are multiplexed together on the VP.The NCA must know the capacity, the number of sources of each type and the Individual CLP requirement!; by each type of traffic.
In [3], an efficient algorithm is developed to compute the minimum capacity required to satisfy all the QOS requirements when this classes of on-off sources are multiplexed onto a single VP.It applies the Markov Modulated Deterministic Process method.
In order to decide: on integratiodsegregation, the different mini" link capacities needed for the multiplexing of the sources of each type must be computed.
We propose the below algorithm to determine the Minimum Capacity required so that all the individual QOS By the integration approach the NCA must be calculated by all the sources of all types and by the most restrictive QOS requirement:

Most Restrictive Integration approach
The Required Capacity calculated with each strategy is taken as a basis for the decision.

DIFFERENT SERVICE REQUIREMENTS BY INTEGRATION APPROACH IN ABSENCE OF PRIORITY BUFFERS
Ifthe network provides priority buffering mechanisms, different Cell Loss Probabilities can be satisfied by each traffic flow.Futhermore, in the integration approach it is not necessary that the most restrictive QOS requirements must be applied to all t r a c classes In absence of priority buffer, it is also possible to provide different Individual Cell Loss Probabilities by using the New Convolution Approach.In this case, it is neither necessary to apply the most restrictive QOS requirements.

CAPACITY REQUIRED
In this case the minimum link capacity required can be calculated with the NCA which allows that different Individual Cell Loss Probabilities to be satisfied by each traf5c flow: Individual CLP Integration approach

INDIVIDUAL CLP
In this section we explain how the New Convolution Approach (NCA) computes the Individual Cell Loss Probabilities.
The NCA evaluates the statistical distribution of the inStantaneous offered rate Y by all established connections on a link, This permits the evaluation of the Probability of Congestion (PC).The PC is a parameter of the link state.It shows the probability that the resources of the link may be surpassed.We focused our study on the bandwidth of the that PC shows the probability that the offered rate Y is larger than the capacity C of the llnk.This is the expression for CLP as shows (l), and this implies that expression (3) is correct.Whenever NCA is used for CAC, individual CLPj can be an upper bound for bandwidth allocation.In this section we investigate the influence of the peak bit rate and the mean bit rate on the capacity required by a VP to support a single trafjic type (figure 2) and the Required Capacity for both integration and segregation approaches (figures 3-5).

TRAFFIC MODEL d
We consider several types of on-off sources modelled by General Modulated Deterministic Process (GMDP).Each traffic types is characterised by the peak rate and by the probability of each state.

(A2O)vs (B 40) vs (C200)
Figure 2 : Required Capacity vs. Individual CLP Fig. 2 shows the Required Capacity varying the QOS requirements (Individual CLP) from 1E-2 to 1 E-1 1, for three types of sources.We consider 2Cb connections of type A, 40 connections of type B and 200 connections of type C.This number of connections bas been chosen in order that the mean load of the system were the same in the three cases.
For the same burtiness a low number of type A connections demand more bandwidth than type B connections.If there are few connections on a link the statistical multiplexing gain may be small.Type C connections demand more bandwidth than type B connections due to the fact that the bustiness of Type C connections is larger than the burstiness of Type B connections.
All type of sources demand more bandwidth whenever the QOS is more stringent.
In figures 3, 4 and 5 we compare the segregation approach wht the integration approach when one type of connections is considered only.Results are plotted variyng the Indwidual CLP (n and m) from 1E-2 to 1 E-1 1.
Figures 3% 4a and 5a show the Required Capacity when using segregation approach and integration approach.
In each figure only one type of connections are considered ("homogeneous traEc").By the segregation one, two VP are considered; each with the same number of connections (NC).By the integration one, there is one VP with double numher of connections (2*NC).Dotted lines show the Required Capacity using the segregation approach and solid lines show the Required Capacity using the integration approach.In figures 'a', the upper line corresponds to n = 11.The horizontal solid lines result from the fact that in the integration approach the Required Capacity is solely determined by the most Stringent QOS requirement.
The load of systeni is the same in all the cases, and it is the double of the considered in figure 2.
Figures 3b, 4b and 5b show the reduction of the Required Capacity when the integration approach is used instead of the segregation approach.In this case, figures b, when m = 2 the upper line corresponds to n = 11, but when m = 11 lit corresponds to n = 2.When the results are greater than 0 the integration approach is better; otherwise the segregation one is better.
These results show that : the integration approach requires less capacity than the segregation approach almost in all cases, when one Individual CLP (n) is fixed, the greatest reduction of the Required Capacity is obtained when the other Individual CLP (m) are the same value, i.e. n = m.By N C 4 O connections of type B, the integration approach requires less capacity than the segregation approach for almost all case:s, except for the cases of n = 2 and m>7, or n>7 and n=2.

Individual CLF'
By NC=20 mnnections of type A, the integration approach requires less capacity than the segregation approach : for almost all cases, except for the cases of n = 2 and m>7, or n>7 and n=2.By NC=20 connections of type A, the integration approach requires less capacity than the segregation approach for almost all cases, except for the cases of n = 2 and m>7 or n>7 and n=2.

HETEROGENEOUS TRAFFIC
In this section we compare the Required Capacity for inkgation and segregation approaches when two type of connections are considered.The load of system is the same i n all the cases, and it is the same that we have considered in the figures below.By N C 4 0 connections of type B and NC=200 connections of type C, the integration approach requires less capacity than the segregation approach for almost all cases, except for the cases of n = 2 and m>6 By N C 4 0 conne:ctions of type B and NC=20 connections of type A, the integration approach requires less capacity than the segregation approach in all cases.
Solid lines are riot completely horizontal because the Required Capacity is plotted for Individual CLP integration approach.The Individual CLP integration approach requires less or same capacity that the Most Restrictive integration approach in ali cases.Therefore, this improvement implies that segregation approach is not the best in some cases.

Figure lb .
Figure lb.Segregation approach.
are guaranteed: integer RquiredCapacity (Number of sources, Required Individual CLP) AGximumCapacity = 1 ; While NCA(Mm"Capacity, Number of sources, Required Individual CLP)='W" r i m Minimmcapacity; Minimumcapacity = + 1; Algorithm to determine the mini" capacity required by a Virtual Path By the segregation approach the mini" link capacity required by a link is the addition of the capacity required by each iype of traflic.Capacity = 0; For i = 1 to Number of SourceTypes Capacity += Requiredcapacity (Number of sources i, Required Individual CLP i); Segregation approach

0Figure 3b :
Figures 3a and 3b, (B 40) + (B 40) vs. (B 80) show that in the segregation approach we consider 40 connections of type B by each VP, where as in the integration approach we consider 80 connections of type B.The Individual CL9 is m for first connections, and n for second connections.

Figure 4a :
Figure 4a: Required Capacity vs. Individual CLP Figure 4b: Reduction of the Required Capacity vs.

Figures
Figures 6a and 6b, (B 40) + (A 200) vs. (I3 40, A 200) show that in the segregation approach we consider 40 connections of type B by one VP and 200 connections of type A by the other, where as in the integration approach we consider in the same VP 40 connections of type B and 200 connections of type A. In this case, the Indwidual CLP is m for type B connections and n for type A connections.